ChatGPT for Hospice Documentation: Risks and HIPAA-Compliant Alternatives

Quick Answers: ChatGPT for Hospice

Is ChatGPT HIPAA compliant for hospice documentation?

No. OpenAI's consumer ChatGPT and the standard API tier do not sign Business Associate Agreements (BAAs) at the consumer level, and OpenAI's data-handling policies historically allowed consumer inputs to be used for model improvement. Using ChatGPT with patient health information (PHI) constitutes a HIPAA breach risk regardless of how careful the individual user is with prompts.

What about ChatGPT Enterprise or Microsoft Copilot?

These tiers offer different data-handling commitments and, in some cases, BAA pathways. But the architecture remains commercial-cloud-based and the security posture does not match FedRAMP High infrastructure. Hospices handling Medicare and Medicaid data, the most regulated category of healthcare data, should use AI infrastructure designed for that data class.

What is the real risk of using ChatGPT with PHI?

Three distinct risks: (1) HIPAA exposure for the hospice as the covered entity; (2) potential use of inputs for model training, with no clear path to delete data already absorbed; (3) reputational and legal risk if a breach is identified during a survey or audit. The risk is highest in clinical drafting workflows where clinicians copy actual visit notes or chart excerpts into the prompt.

What is a HIPAA-compliant alternative?

A generative AI platform that (a) is hosted on FedRAMP High infrastructure (AWS GovCloud is the only commercially available FedRAMP High GenAI platform), (b) includes a Business Associate Agreement at every tier, (c) never uses customer inputs to train the model, and (d) is built for the document sizes and clinical reasoning needs of hospice work. Hathr.AI is the only platform that meets all four criteria.

Why Hospice Clinicians Reach for ChatGPT

Documentation burden is consistently cited as a top driver of hospice clinician burnout. Recertification narratives, IDG summaries, eligibility justifications, and HOPE documentation eat into evenings and weekends. A free, easy-to-access AI tool feels like a way to claw back time — and in many other industries it is.

Hospice clinicians are not the problem. The architecture they are reaching for is.

The Three Core Problems with Consumer AI for Hospice

1. No BAA at the consumer level

The Health Insurance Portability and Accountability Act (HIPAA) requires that any third party with access to Protected Health Information sign a Business Associate Agreement with the covered entity. Consumer ChatGPT and consumer Gemini and consumer Claude do not sign BAAs with individual subscribers. A clinician pasting a visit note into a personal ChatGPT account has shared PHI with a third party that has not signed a BAA. That is a HIPAA breach.

2. Training on inputs

OpenAI's terms have evolved, and some tiers now offer training opt-outs. But for consumer ChatGPT and many API workflows, inputs may be used to improve the model. Hospice data — even de-identified data — entering a global training pipeline is not what any hospice consented to when the patient signed the election statement.

3. Infrastructure security posture

Hospice handles Medicare and Medicaid data, the most regulated category of healthcare data. The Department of Health and Human Services hosts its sensitive workloads on AWS GovCloud, a FedRAMP High environment. Hospice clinical and operational AI should match that posture, not run on commercial-tier cloud that is appropriate for consumer chat but not for federal health data.

The McKinsey Lilli Breach: A Warning Shot

In a widely reported incident, CodeWall's offensive AI agent compromised McKinsey's internal AI platform Lilli in under two hours using SQL injection — one of the oldest vulnerability classes in software. The breach exposed 46.5 million chat messages, 728,000 files, and 57,000 user accounts. The lesson for hospice: internally-built AI on commercial-tier infrastructure inherits all of the security risks of commercial-tier infrastructure. FedRAMP High exists precisely because federal data needs a higher bar.

The Hathr.AI Alternative

Hathr.AI is the only generative AI platform built on AWS GovCloud, the FedRAMP High environment used by the Department of Health and Human Services. Key differences from consumer AI:

• BAA included with every plan. Signed in 24 hours, not months.
• Zero data retention for training. Hathr.AI never uses customer inputs to train the underlying model.
• FedRAMP High infrastructure. The same environment HHS uses for its sensitive workloads.
• 500,000+ word document handling. Five times more than commercial Claude or GPT, which is what full hospice chart reviews require.
• Built for healthcare workflows. Documentation, summarization, chart review, IDG summaries, recertification narratives — not generic chat.

What Hospice Clinicians Should Do This Week

• Audit current shadow AI use. Survey the clinical team anonymously. Ask whether anyone has used ChatGPT, Gemini, or Claude.ai for hospice work. The answer is almost always yes.
• Adopt a sanctioned alternative. Make a HIPAA-compliant AI available so clinicians do not have to choose between burnout and breach risk.
• Update the AI use policy. Many hospice compliance manuals predate generative AI entirely. Add explicit language.
• Train on prompt hygiene. Even on a HIPAA-compliant platform, clinicians should know how to prompt effectively. Hathr.AI publishes a prompt library specifically for healthcare.

Frequently Asked Questions

What if a clinician used ChatGPT for a hospice note once — is that a breach?

Consult your compliance officer and HIPAA counsel. If PHI was shared with a non-BAA third party, the incident likely meets the definition of an impermissible disclosure under the HIPAA Privacy Rule. Whether it rises to a reportable breach depends on the specific facts and risk analysis required by 45 CFR 164.402.

Is the OpenAI "Team" or "Enterprise" tier safe for hospice?

The Team and Enterprise tiers offer stronger data-handling commitments and BAA pathways may be available. However, infrastructure remains commercial cloud (not FedRAMP High), and the architecture was not built for federal health data. For Medicare and Medicaid data, that gap matters.

Why does FedRAMP High specifically matter?

FedRAMP High is the federal certification level for systems handling the most sensitive non-classified federal data, including most health data. The Department of Health and Human Services operates its sensitive workloads in FedRAMP High environments. A hospice AI that runs on FedRAMP High infrastructure inherits the security controls federal agencies require.

Can our hospice build its own ChatGPT wrapper to make it compliant?

Building a HIPAA-compliant wrapper around a consumer model is technically possible but operationally expensive, requires ongoing security maintenance, and still depends on the underlying model vendor's data-handling commitments. Most hospices are better served by adopting a purpose-built compliant platform.

Sources and Further Reading

• HHS HIPAA Privacy Rule
• FedRAMP program overview
• McKinsey Lilli AI platform breach (Hathr.AI analysis)
• Why Hathr.AI

This article is for informational purposes only and does not constitute legal, clinical, or compliance advice. Hospices should consult HIPAA counsel before making changes to their AI use policy. Last reviewed: May 2026.

Give your clinicians a HIPAA-compliant AI before they reach for one that isn't. Start a 7-day free trial of Hathr.AI — the only generative AI built on AWS GovCloud, with a signed BAA included on every plan.